Content area
Abstract
Learning from Demonstration (or imitation learning) studies how computational and robotic agents can learn to perform complex tasks by observing humans. This closely resembles the way humans learn. For example, children can imitate adults in a variety of domestic tasks, such as pouring milk from a carton to a cup or cleaning a table, and after one or a few observations, they can replicate the action with relative accuracy. This research tests the feasibility of a virtual agent that can process multimodal (linguistic and visual) captures of actions and learn to reenact them in a simulated environment.
In particular, this thesis looks into the problem from two perspectives. The first perspective is taken in a realistic setup, in which we will teach learning agents skills by showing real demonstrations of the skills. We will only focus on a small set of actions involving spatial primitives. The objective is for agents to learn to perform complex action skills from a limited amount of training samples. We will also try to teach agents in an interactive environment, in which immediate feedback is provided to agents to correct them in due time.
The second perspective is taken in a synthetic setup. In this setup, a parallel corpus mapping natural language instructions to demonstrations is generated by soliciting human descriptions for complex action demonstrations in a two-dimensional simulator. The problem is framed in a sequence to sequence translation framework. We will discuss the difference between the two perspectives, as well as the advantages and disadvantages of corresponding learning frameworks.
The system is composed of the following components: a. an event capture annotation tool that captures human interaction with objects using Kinect ® sensor; b. an event representation learning method using recurrent neural networks with qualitative spatial representation feature extraction; c. action reenacting algorithms using reinforcement learning and sequence to sequence methods; d. evaluation methods using 2-D and 3-D simulators.
